RuntimeError: einsum(): the number of subscripts in the equation (3) does not match the number of dimensions (4) for operand 0 and no ellipsis was given about deq

[jerrybai1995]

Hi @QiyaoWei ,

Thanks for your interest in our work. It seems like you are using a 4D tensor rather than a 3D tensor. You should do the following instead:

bsz, C = 128, 3
H, W = 32, 32
test = torch.rand(128, 3, 32, 32).to(device)
layer = ResNetLayer().to(device)
f = lambda x: layer(x.view(bsz, C, H, W)).view(bsz, C, -1)
zz = broyden(f, torch.zeros_like(test).view(bsz, C, -1), threshold=100, eps=1e-1)['result']

As a side note, I suggest adding some normalization layers to your ResNetLayer design. Without normalizations it is likely that this layer will not converge with random initialization.